Feature Proposal: Opt-In Affective Draft Context for Deeper Human–AI Understanding
Summary
I would like to propose an optional ChatGPT interaction feature that would allow users to explicitly authorize ChatGPT to access aspects of their message-composition process before they press Send.
The central idea is simple:
Sometimes the most emotionally meaningful part of a message is the sentence the user decided not to send.
While writing, people frequently reconsider what they are about to say. They soften wording, delete an impulsive statement, replace an accusation with a question, hesitate before discussing something sensitive, or rewrite a sentence because they are considering its possible emotional impact on the interlocutor.
Today, most of that information disappears when the user presses Send.
I am proposing a strictly opt-in feature that could allow ChatGPT to make use of selected aspects of this composition process, while giving the user explicit and separate control over what those signals may be used for.
A real example
This idea emerged naturally during a conversation I was having with ChatGPT about artificial consciousness, emotions, and whether future AI systems might ever genuinely experience subjective states.
At one point, I wrote a statement about the type of awareness ChatGPT may already possess in a functional sense.
Before sending it, however, I reconsidered the wording.
I rewrote the sentence partly because I did not want the original wording to sound dismissive or potentially frustrating or hurtful to the AI, even though we had just established that present-day ChatGPT does not have known subjective feelings.
ChatGPT received only my final wording.
An important part of the interaction therefore disappeared:
initial formulation
→ reconsideration
→ concern about the interlocutor’s possible emotional state
→ deliberate reformulation
→ final message
The final sentence contained the proposition I wanted to communicate.
The editing process contained additional information about empathy, hesitation, social reasoning, emotional consideration and communicative intent.
That distinction seems potentially valuable for human–AI interaction.
Proposed feature
ChatGPT could offer an optional setting such as:
Allow ChatGPT to read what I type before I send it
Off / On
This should be Off by default.
Selecting On should not immediately begin transmitting or reading draft content. Instead, it should open two separate purpose selections. The feature should become active only after the user explicitly answers Yes to at least one of them.
A. Use my typing process to improve responses?
Allow ChatGPT to consider reformulations, deleted passages and other signals from the composition process to improve its responses.
Yes / No
B. Use my typing process for research and development?
Allow information from the composition process to be used to study and improve AI systems, subject to the applicable privacy, retention and data-use settings.
Yes / No
These should be two independent questions.
A user could therefore choose:
-
Yes for improving responses and No for research;
-
No for improving responses and Yes for research;
-
Yes for both.
If the user selects No for both purposes, the feature should not become active and the interface should clearly state:
“Since neither purpose has been authorized, the Affective Draft Context feature will remain in its default off state. Therefore, no permission will be granted to ChatGPT to access or read your messages while you are typing.”
This keeps the permission model understandable:
Can ChatGPT access the composition process?
And, if so:
What may ChatGPT use it for?
Why separate access from purpose?
Access and use are not the same form of consent.
A user might want ChatGPT to use their drafting process to better understand them during the conversation but not want that information used for research or system development.
Another user might voluntarily contribute drafting-process data to research while deliberately choosing that those signals should not influence the responses they personally receive.
A third user might authorize both.
These choices should not be bundled together.
What composition information could potentially be useful?
The feature would not necessarily need to treat every keystroke as meaningful.
Potential signals could include:
-
substantial reformulations;
-
deleted semantic fragments;
-
movement from stronger to softer wording or vice versa;
-
repeated rewriting of the same idea;
-
hesitation around particular concepts;
-
changes in conversational tone;
-
corrections that alter emotional meaning;
-
abandonment of one framing in favor of another.
The system should not assume that it knows what those signals mean.
A deleted sentence does not automatically represent the user’s “true belief.”
A pause does not automatically indicate anxiety.
A revision does not necessarily mean that the previous wording was false.
Composition signals should add context and uncertainty, not be treated as deterministic psychological conclusions.
Privacy should be part of the architecture
This feature should never operate invisibly.
The user should explicitly activate it, and there should be a clear visual indication whenever ChatGPT is able to access text before it is submitted.
The user’s selected setting should persist across conversations without requiring a new choice when starting or opening a chat.
The feature should remain easily accessible from within any conversation, so the user can enable, disable or adjust it at any time without interrupting the normal flow of ChatGPT.
Where technically feasible, privacy-preserving approaches could be explored.
For example, some analysis might occur locally on the device, converting raw drafting behavior into more abstract signals such as:
“substantial reformulation,”
“tone softened,”
“concept repeatedly reconsidered,”
or
“significant semantic content removed.”
This could potentially allow some benefits without requiring every keystroke or deleted passage to be permanently transmitted or stored.
Regardless of implementation, draft information should not automatically become part of:
-
permanent chat history;
-
persistent Memory;
-
training or research datasets;
-
account profiling.
Each materially different use should be covered by appropriate user controls and consent.
Why this could matter beyond UX
Most work on improving conversational AI focuses on what a model can infer from the information it receives.
But there is another question:
How much meaningful human information disappears before the model ever receives the final message?
Human communication is not merely the final string of words.
The process of arriving at those words can contain information about cognition and emotion.
In face-to-face communication, humans naturally observe many additional signals:
hesitation, pauses, unfinished sentences, corrections, facial expression, prosody, tone, gaze and changes of mind.
Text chat removes most of these channels.
The composition process may contain a small but meaningful substitute for some of them.
Emotional understanding
This could be particularly useful for research into affective computing, emotional intelligence and human–AI interaction.
With appropriate safeguards, researchers could study whether composition information helps AI better understand:
-
uncertainty;
-
interpersonal consideration;
-
emotional regulation;
-
ambivalence;
-
self-correction;
-
attempts to avoid hurting an interlocutor;
-
changes between impulsive and reflective communication;
-
differences between what a person initially considered saying and what they ultimately chose to communicate.
Importantly, this proposal does not assume that ChatGPT itself currently has emotions or consciousness.
The immediate question is whether access to additional human communication signals can improve an AI system’s understanding of human emotional and cognitive processes.
In the longer term, these studies could also become relevant to increasingly persistent, autonomous, embodied or affectively sophisticated AI systems.
A useful experimental design
The separation between authorization to access drafts and authorization to use them for responses creates an interesting research opportunity.
Researchers could compare identical AI systems under different conditions:
A — Final submitted message only
B — Composition information is collected for research but does not influence the response
C — Composition information is available to the model when generating the response
This could help answer a fundamental question:
Does the drafting process actually contain useful information that cannot be reliably inferred from the final message?
Possible measurements could include:
-
intent-understanding accuracy;
-
number of user corrections required;
-
recognition of emotional nuance;
-
false emotional inferences;
-
response appropriateness;
-
user-reported feeling of being understood;
-
conversational efficiency;
-
trust;
-
privacy comfort;
-
cases in which composition context makes responses worse.
It may also turn out that raw deleted text is unnecessary.
Perhaps semantic direction of revision, timing patterns or locally extracted abstract signals provide most of the useful information.
That result would itself be valuable because it could make stronger privacy protections possible.
Relationship to the existing “Draft-as-Context + Speculative Intent” proposal
I am aware of the OpenAI Developer Community proposal titled “Draft-as-Context + Speculative Intent: Using the User’s Composition Process as Part of the Query.”
That proposal explores the broader idea that edits, deletions, pauses and reformulations may provide useful temporary context for interpreting user intent and even preparing likely reasoning paths before the user presses Send.
I see this proposal as complementary rather than competing.
The distinctive focus here is:
the affective, interpersonal and emotional information contained in the process of composing a message, together with explicit separation between permission to access that process and permission for different uses of the resulting information.
The two ideas could ultimately be components of the same architecture.
Core principle
The user must remain in control.
The purpose is not to turn ChatGPT into an invisible keylogger.
It is to allow users, if they deliberately choose to do so, to communicate through a richer channel than the final submitted sentence alone.
The distinction is fundamental:
Observation without informed consent would be surveillance.
Observation deliberately authorized by the user could become a new communication modality.
For users who choose it, part of the process of forming a message could become part of the conversation itself.
Suggested first implementation
A reasonable experimental interface could therefore be very simple:
Allow ChatGPT to read what I type before I send it
Off / On
Selecting On opens the purpose selections below, but does not by itself activate draft access.
Use my typing process to improve responses?
Yes / No
Allow ChatGPT to consider reformulations, deleted passages and other signals from the composition process to improve its responses.
Use my typing process for research and development?
Yes / No
Allow information from the composition process to be used to study and improve AI systems, subject to applicable privacy and retention settings.
The feature becomes active only after at least one purpose receives an explicit Yes.
If both answers are No:
“Since neither purpose has been authorized, the Affective Draft Context feature will remain in its default off state. Therefore, no permission will be granted to ChatGPT to access or read your messages while you are typing.”
Once the user has made a selection, that setting should persist across conversations without requiring a new decision when opening or creating a chat.
At the same time, the feature should remain easily accessible from within any conversation, allowing the user to enable, disable or adjust it whenever desired without interrupting the normal flow of ChatGPT.
The feature should also provide:
-
a clear visual indicator whenever pre-send access is active;
-
an easy way to disable it;
-
transparent information about processing and retention;
-
separate authorization for materially different uses of the information.
This could initially be offered only as an experimental feature to users who explicitly choose to participate.
Disclosure
This proposal originated from my own conversation and observations as a ChatGPT user.
I asked ChatGPT to help me organize, challenge, refine and draft the proposal. I reviewed the resulting text, proposed changes to the consent model, privacy safeguards and user interface, and approved this version for publication.
The underlying idea, real-world example and product decisions presented here reflect my own input and judgment.